大豆品种快速准确的鉴别, 对于鉴定种子品质、 净化种业市场以及保障粮食安全具有重要意义。 为解决传统农作物品种鉴别方法中存在精度差和效率低等问题, 采用拉曼光谱结合特征波长提取方法建立偏最小二乘(PLS)鉴别模型, 对黑龙江省4个高蛋白大豆品种（黑农88、 黑农98、 绥农71以及绥农76）进行快速鉴别。 随机蛙跳(RF)算法是一种通过迭代计算变量被选概率, 以确定变量重要性的新型特征波长选择算法, 可以有效剔除全光谱数据中的冗余信息。 该方法存在初始变量集随机性、 所需迭代次数大、 阈值选取不确定的问题, 因此提出一种基于最小绝对收敛与选择算子(LASSO)回归的改进随机蛙跳（MRF）算法。 采用LASSO算法提取与属性变量最相关的特征波长点作为RF初始变量集F0, 消除初始变量的随机性, 在此基础上开始迭代计算, 可以减少无用迭代次数, 提高模型的预测精确度。 RF算法通过设定阈值的方法选择变量, 因此提取的特征波长往往具有不确定性。 改进如下: 首先去除被选概率为0的变量, 对于排序后变量以10个波长点为间隔, 每次增加1个间隔建立特征波长与大豆品种属性的偏最小二乘回归模型, 当交叉验证均方根误差(RMSECV)取最小值时的建模波长为优选特征波长。 以MRF优选特征波长作为输入变量建立PLS鉴别模型, 并与全光谱以及常用的RF、 LASSO和ElasticNet特征波长选择算法建模结果进行对比分析。 结果表明, MRF算法提取300个特征波长点, 仅占全谱波长的9.37%, 有效筛选了关键特征变量, 简化了模型复杂度。 预测结果中均方根误差(RMSEP)和决定系数(R2p)分别为0.246 9和0.951 2, 识别准确率达到100%, 为所有模型中最优。 拉曼光谱结合MRF算法可以实现大豆品种的快速鉴别, 同时也为其他农作物品种的快速鉴别提供了一种新思路。

Rapid and accurate identification of soybean varieties play an important role for identifying seed quality, purifying the seed market and ensuring food security. The traditional identification methods of crop varieties have the problems of poor accuracy and low efficiency. Therefore a PLS identification model was established by Raman spectroscopy combined with characteristic wavelength extraction to fast identify four high-oil soybean varieties (Heinong 87, Heinong 89, Suinong 38 and Suinong 77) in Heilongjiang Province. RF is a new characteristic wavelength selection algorithm that determines the importance of variables by iteratively calculating the selected probability, which can remove redundant information to a great extent in the full spectrum. However, this method has the disadvantages of the random initial variable set, a large number of iterations and uncertain threshold selection. Therefore, an improved random frog (MRF) algorithm based on LASSO regression was proposed. In order to get rid of the randomness of the initial variable set in the RF algorithm, LASSO was used to extract the characteristic wavelength point most related to the attribute variable as an initial variable set F0. On this basis, iterative calculations were carried out to reduce the number of useless iterations and improve the models prediction accuracy. In addition, RF selects variables by setting a threshold, which leads to the uncertainty of the extracted characteristic wavelength. The improvements were as follows: Firstly, the variables with the selected probability of 0 were removed, taking 10 wavelength points as intervals for the sorted variables. Then, the partial least squares discriminant analysis model between the characteristic wavelengths and soybean varieties was built by adding one interval each time, and taking the wavelength subset with the smallest RMSECV as the selected characteristic wavelengths. The PLS-DA model was established with the selected characteristic wavelengths of MRF as the input variables and compared the prediction performance with full spectrum and other characteristic wavelength selection methods of RF, LASSO and ElasticNet algorithms. The results indicated that the MRF algorithm selected 300 characteristic wavelength points, accounting for only 9.37% of the full spectrum, which effectively screened the key characteristic variables and simplified the complexity of the model. The RMSEP and R2p were 0.246 9 and 0.951 2 respectively, and the identification accuracy reached 100%, which was the best among all models. Therefore, Raman spectroscopy combined with MRF algorithm could achieve the fast identification of soybean varieties and provide a new technique for the fast identification of other crop varieties.